Computing Optimised Parallel Speeded-Up Robust Features (P-SURF) on Multi-Core Processors

Abstract

This article presents a novel CPU-based parallel algorithm (P-SURF) that computes the Speeded-Up Robust Features (SURF), a local descriptor that is able to find point correspondences between images in spite of scaling and rotation. The algorithm presented here parallelises all the seven major steps found in the original serial computation. The task in each of the steps is decomposed and the fractions are assigned to running threads bound onto distinctive processors. The implementation of the algorithm was tested using randomly selected images in regard to performance, scalability and stability. The results showed that its performance on mid-level Intel Core Duo processors was comparable to that of some fast GPU-based SURF implementations. For example, on a testing system equipped with an Intel Core Duo P8600 at 2.4 GHz, P-SURF was able to extract and represent features from a 640 × 480 image at a rate of 33 frames per second. The experimental results also revealed that, instead of leaving the threads to the kernel for processor assignment, assigning hard processor affinity by the algorithm produced better performance and stability.

Towards the Optimal Hardware Architecture for Computer Vision

“…1 shows performance comparison of a computationally intensive algorithm as SURF (Bay et al, 2006) for different platforms. PCs offer uneven performance if heavily use multithreading programming (Zhang, 2010) or a straightforward implementation (Bouris et al, 2010). GPUs feature very large performance at the expense of high power consumption.…”

“…Performance Power (W) Bouris et al (2010) Intel Core 2 Duo 2.4 GHz < 7 fps N/S Zhang (2010) Intel Core 2 Duo P8600 2.4 GHz 33 fps 25 Terriberry et al (2008) nVidia GeForce 880 GTX 56 fps 200 Bouris et al (2010) Xilinx Virtex 5XC5VFX130T 70 fps < 20 Table 4. On-chip retinal vessel-tree extraction (hardware-oriented algorithm) on different platforms.…”

Video Key Frame Detection Based on SURF Algorithm

“…SURF is based on SIFT, and it uses Integral Images instead of DOG (Difference of Gaussian), which allows to work much faster than SIFT [24] [25]. It can also be accelerated by a GPU [38] and it has a parallel implementation [42]. SURF is based on image keypoints (interesting points), which allows to extract local features from an image.…”

Adaptive Pipeline Parallelism for Image Feature Extraction Algorithms

“…There have been some researches on parallelizing these algorithms, such as [16][20] [21] [22] [23]. However, no satisfied results are achieved in these works.…”